 Good afternoon, brilliant nerds, and welcome back to the Mile High City. We're here at Supercomputing 2023. My name is Savannah Peterson, and you're watching theCUBE. I'm joined by the fabulous Dave Nicholson, and we've also got a fabulous fellow analyst with us, Brian, welcome to the show. Thank you, fabulous. That's the nicest thing I've been called in weeks. That's because we were just getting to know each other. Brian, no, I'm kidding. Give it time, give it time, give it time. So Brian Bieler, you're with storagereview.com. Obviously, an event with a lot of storage as front of mind. What's the number one conversation or thing you're seeing day one here walking around the show floor? Well, I mean, the best thing about Supercomput is the stuff. This is the show. It is a physical show, for sure. This is the most stuff show of the calendar year. I mean, you guys go to all these trade shows. You've been to dozens of them this year as well, I'm sure. And so many are focused on high-level messaging on maybe some solution stuff or as a service, but this is very much stuff-oriented. And so I realize the Supercomput audience is like, you've called them, what'd you call them, creative nerds? Yeah. Okay, so half are creative software nerds. The other half are creative hardware nerds, and here we jumble them all up. But you can talk to somebody about servers and GPUs and they'll have no idea, because I've never seen it before, because they're software nerds and the hardware nerds don't really necessarily know about all the software. So this is a great amalgamation of everybody, and it's such a global audience. I have fun learning just what, talking to a guy last night, waiting to get in. What's going on in India? I mean, that's where you are, what are you thinking about? Because it's so different than what we sort of get myopically focused on in my world anyway sometimes. So I think it's a great reset. I'm really glad you brought up the global nature of that. We had someone from Dell who lives in Romania on the show. We also had some Belgian guests earlier. There's a whole Korean pavilion behind me. Oh, absolutely. Japan has a sticker activation going on for boosting around the show floor. We talk a lot, big key theme here is the democratization of AI, and I feel like sometimes that's a real big buzzword, especially in the Valley. Coming to shows like this give me hope, and I'm curious if you guys agree, because I do see small startups, large enterprise, academics, government, and countries from all over the world collaborating to help build these future solutions. David, would you agree that's the good stuff? The magic? Or am I just drinking my own community Kool-Aid? I thought my role here was to always disagree with you, Savannah. We can, I mean. No, no. You know I don't mind a little confrontation. Yeah, no, no. No, I agree completely. It's interesting. I mean, it is absolutely a global gold rush in a lot of ways. Yeah. Number one thing, you know, when asked about the hype cycle, the answer is yes, the hype is a hype cycle, but it's real. I think that the evolution that AI will bring is real. You mentioned folks in coming from the hardware perspective or the software perspective, you focused on storage, right? For a period of time. How has that changed? And do people appreciate it? I mean, just anecdotally, I can tell you I've had conversations with people who struggle over this idea of creating a quote unquote persistence layer like it's never been done before. Right. What are your thoughts on the shifting sands of storage outside of the array, inside of the array, in the server, out of the server? I mean, there's a million ways to go with that. And just as a point of clarity, when you say in the valley, I assume you mean the Ohio River Valley? Or are you talking? That's exactly what I'm referring to. I know you're a since boy and, you know, yeah. So in the Ohio River Valley, we are very concerned about democratizing AI. So I'll give you that one callback, but I do think it's, I didn't really know that that was a theme, but it's something that we've been talking about for a long time is that there's all these tools available, whether it's storage, GPUs, compute, any of these things, but they're not equally accessible, I would argue, to organizations of all size. So we're here looking at the HPC audience today, but there's a lot of enterprise customers here too, looking to see what these guys are doing at the bleeding edge and how that will translate into enterprise solutions in the data center. Dell's here, we spent some time in their booth. We've got a number of GPU servers that are really great, but it's not just so easy as placing the order and throwing this in your rack. There's all sorts of new things. And to your point about storage, networking and storage are a big part of that because these GPU servers specifically are not storage heavy. They may have a drive per card in their configuration. So how do we fuel these GPUs? Because I'm going to drop a million or a couple million dollars and from an enterprise perspective into an enterprise data center to use AI to help research or improve our businesses or whatever we used to call business information or analytics three years ago. All the infrastructure's got to be there to support it. You said, are there changes? I mean, yeah, there are changes everywhere, but we're in this period that feels so fragmented. I don't know that there's one right answer because we can look at the banners scattered around here. You've got software-defined object storage that'll run on anything and it's cheap and deep. That's one thing. You've got these big parallel file systems, there's two over there, that's another thing. You've got the standard file unstructured players that have been after this market for decades. So it's wild. I don't know that there's any one thing happening. There's so many things happening which does make it even more challenging to sift through what's appropriate for your organization or your deployment. So in the world that is storage review, what I just heard you say is lots of job security. Because- Perpetual, yes. Perpetual job security, perpetual movement in change. Where do you think these enterprise instantiations are going to occur? Do you think people will default toward an older model of deploying things internally on-premises or does the sort of fear of missing out that people experience now, is that the final push where it's like, okay, no, by default, this is going to be SaaS. We want this as a service delivered by, whether it's a hyperscale cloud provider or some niche service provider. Is this yet another nudge in the direction of cloud or is it a validation of hybrid? Any thoughts? I mean, I think at the lowest level, the argument should start with your organization should be doing AI. So that's like foundational thing because if you're not involved, and I don't even think this is a Fortune 500 problem, I think this goes all the way down the stack. And depending on where you are in that stack, I think answers the question in terms of where you should be. I mean, you talked about democratization. I think this is so important because I think there's a ton of gatekeeping in this industry that I find really frustrating that professionals that are in here that have been doing this for years will look at these infrastructures or storage or GPUs or servers and say, well, this is the way to do it with this software stack and this is the output and here's what you get. And if you're not doing it that way, you're doing it wrong. And I think that's a, I mean, it's a little sickening, honestly, because- I agree with you, it grosses me out. Because these organizations that are trying to do it and they're starting with a workstation, with a couple A6000s in it, you're still talking about a $50,000, $60,000 investment in hardware and they're learning and they're trying to do it and they're trying to get better and trying to figure it out. This is not AI, that's still AI. It's still productive and this workload that can start on a workstation, I think the question is, where does it go next? Does it go to on-prem infrastructure? Does it go into the cloud? Does it do both? And part of it's, I think about cost and availability. We were told just a week and a half ago that the wait for H100 systems, you place your order today, is somewhere around 40 weeks. And so NVIDIA's doing the pivot. That's pretty wild. Well, and the wait is for who? Not me getting one delivered to my house. There are very large organizations that are at the head of the line. Oh, absolutely, and they will command attention. Yes, so the average enterprise waiting for five of them versus someone who is buying as many as can come off of the assembly line, as can be produced. So that's perhaps a driver in terms of the where question. Is can I get access to the hardware? And everyone needs H100 and NVIDIA says, well look, we've got this shiny L40S, it's available now, relatively speaking, a couple weeks. And it's opening up the market I think too for AMD, Intel, and others with accelerators that come in with perhaps lower cost options that may not be quite so powerful, but the AI you can do today is better than the AI what you can do waiting for your GPUs to show up. Well, I mean, if we're talking about 40 weeks, that's a whole year of lost development, essentially. And there are a lot of cars that can get you over the finish line right now for a lot of different types of processing. Get you something, right? Which is I think is better than nothing. But the cloud, of course, is very much a player too. As you said, we recently did some work with OVH cloud who's got what would be considered by this audience sort of lower end V100s in the cloud. But again, I really think it's about accessibility and what can you get and affordability and how can you measure your investments in AI? If you're starting, it's a whole different set of metrics than if you're already succeeding and already have AI teams that can go in and leverage an Oracle cloud bare metal eight by H100. But that's not free either. I mean, that's a year commitment at a minimum and a lot of cycles that you've got to burn. So you've got to be ready to do it and to your point about data, you've got to be able to move the data to these GPUs. That's a whole nother issue with cloud based GPU computing. It really is, I think. And just for anyone listening, I thought this was cool. We had Johnny Dallas, the founder of ZEAT on last night and he's created a website, gpucost.com, that shows you the availability, the pricing and the providers of every one of these GPUs. And I've been scrolling through between some of our breaks and it is wild like you're talking about to see the delays that we're seeing now and anticipate what may happen as supply chain challenges become more complex and demand becomes higher and moving forward. Having a strategy about your hardware right now, I think is absolutely imperative. Building on that, I just think it's an awesome point and we haven't touched on it yet. And not as busy as it was last year because I think it's getting overshadowed by this very large AI presence in the room is quantum and hybrid quantum. And when we're talking about processing speed, how are the amount of compute we're going to need there? How do you think that's going to affect the market? It's just another tremendous shift, right? I think that's another strategy where maybe not today, right? But I think a lot of organizations, whether they're these big HPC installments or enterprise are again have to be evaluating it to see where that puck is moving to try to have a strategy there. And I think if this GPU availability has taught us anything, it's that the planning is absolutely critical and it's not like the old days of pre-COVID where you could walk into your DISTY and buy whatever you wanted to buy and it'd be on your doorstep in a couple of days or weeks at worst, much more strategic. And there's so many more choices and quantum's another one where it's like, what am I trying to do with this? And really understanding what the output is that you're going after, what the benefits are of the technology and having a plan to consume that, again, either on-prem or in the cloud or wherever. I mean, you didn't bring up edge, that's the other big, that's the other. I was going to drive there, because yeah. Yeah, I mean, that's another one and another one where storage is making a big impact. We've got a project underway right now actually where we're developing our own LLM for scientific research. So Jordan, yeah, it's pretty fun. We're going to open source it too. But Jordan, who's on my team has a passion for astrophotography. And so when you rig up these high-res cameras to a telescope and you're out in the middle of nowhere, you've got all sorts of thermal issues, you've got power issues because you're going to have to use battery most likely to power these things. And you've got data issues because these guys do not want to get rid of any image they've ever taken because as the model gets better, what it does at a very high level is it takes all of these images, it runs through an inferencing engine at the edge or can run at the edge on a smaller GPU and clean up the images and then send back the final image to your data center for more deep processing on something A100, H100, whatever. This whole data modality is tremendously punishing because you've got all this data, you've got Starlink is your best in class communication at the far edge. And you've got thermals and humidity and snow or dust and all these other things that are really terrible. So we've run this rig in Arizona, it's going, we've got a new one running on Dell XR servers with these super dense, solid-ime SSDs that now go up to 61.44 terabytes in a single U.2, which is amazing. So a server that has four bays can now support whatever that's 61 times four, the big math, I know. And we can really do all that data collection but we're talking about data mobility to the cloud. We're looking at something where SneakerNet is still the fastest way to transport that data. And as much as we're all excited I'm excited about Flash and the high-end technologies here. There's two guys over there with tape libraries that are still fundamentally where a lot of this stuff goes to reside that none of us want to get rid of and tape, like, Flash is cool and all but tape's sort of still pretty badass. I don't know, can we say badass on this show? We can now. I give you permission. Too late, too late, live. My apologies. Don't apologize, it's all right. Well, John's flexible as long as it's the right audience. We're not gonna find anyone. We're giving our hot take, sometimes includes hot language. There we go. So there's absolutely nothing wrong with that. I was thinking about it. I saw, and this is the first time I've seen this, there's a couple booths giving away GPUs which is, first of all, that's how you know you're at supercomputing. But second of all, in the current state of things I thought they should have somebody guarding that depending on how nice it was. I would say so. I mean, you're watching the tracker. I mean, in each 100 these days is what, 30K retail or something like that? Yeah, men. Yeah, I mean, even the L4s which we've been using the A2s for the inferencing which is like the easy GPU solution. I think those guys in retail volume are like 24, 2300 bucks. I mean, it's not nothing. This is a chunk of change that we are talking about. Yeah, I'm, I'd like to get your take on this. I'll throw this out there. Okay. You can tell me what you think. I'm curious about the utilization of all of the GPUs that have been installed. Absolutely. Out, especially in hyperscale cloud providers. There is a race to make sure that they don't miss out. Along the way that those supply chain issues, if you will, that arise, get people to take a closer look at what you can do without those GPUs. So training and inference aren't the same thing. Inference doesn't have to have, inference when you're typing into an engine that is checking your spelling needs to happen right away. Inference doesn't always have to happen instantly. And so there are massive cost breaks when you move away from using dedicated, most expensive hardware down to other tiers. And so I see a battle shaping up among these sort of Nvidia ecosystem versus the, hey, wait, you can use more generic things, more cost effectively, number one. And number two, it'll be very interesting to see if in fact all of the people gobbling these GPUs right now are going to be able to effectively monetize them. You said first everybody should be doing AI. I'm sure you would agree that what you meant was first decide what the business problem is you're trying to solve or what the big, it goes without saying. Then you decide, are these tools something that I can work with? Feels like the hype cycle is driving a lot of activity which is great for all of us. It'll be interesting to see utilization wise. Do you have any insight into what that looks like now? I wish, it would be great if the clouds would be a hair more transparent about what is being used because that might actually influence some other people's decisions and some consumption models that may be important for supply chain as you go. Is anyone waiting to spin up a cluster of GPUs that are maybe virtualized with VMware running in AWS? Are people waiting? Is there a three month waiting list to get in? We're talking about 40 weeks. You can have a baby in 40 weeks but you can't get a GPU? It's kind of. I mean, I haven't compared it to baby delivery. I was going to say that was a really bold choice. That's a hard thing, I like it. I have four kids, everything gets compared to babies at some point. Mine are too old, I can't remember the whole process but I'll take your number as fact. I mean, I think if you walked into any one of these clouds and said I'm ready to sign a contract for a couple years for these eight way systems I don't think you have a big problem gaining access to it. In the immediacy of the cloud has always been the historical benefit of the cloud, right? But utilization, I mean, I think you even have to take that back to the workstations where we started at kind of the entry AI model or the lower cost influencing cards. It's like, if you're going to make the 50, 60, $100,000 investment in these pieces of equipment you want to make sure that that's up all the time. So we've got some great systems in from the major workstation providers that are here and one of the arguments we've been making lately is put these things in the data center and figure out how to intelligently provision them and share them with your team so that when you need one in your valley which is a different time zone than my valley we can get more utilization out of the machine and have better return on that hardware investment. But I think it's true through the whole stack. No one wants to drop a couple million dollars on a big time GPU stack from any of these vendors. We've been playing a lot with like the 9680 from Dell. Those things are absolutely amazing. It must be used all the time or you're wasting or at least not maximizing your investment on these things. So there's a lot of that to be said and a lot of that to be figured out still honestly in terms of how we keep these things pushing all the time and get those insights and then of course ask the right questions, as you said. That's such a good point. We had Grock with a Q, not with a K, not Elon's Grock on the show just before you and it was actually, we were talking about batches and how small batches can actually be incredibly expensive to run it to your point, not optimized and having inference tools like Grock, like other players in the game to help figure out when to run and how to optimize that makes the hardware worth the investment or not. I mean, to your point, I think it's a really interesting trade-off between... Plus they had a llama outside this morning. They did. I heard about that, which I still contend that it could be a wild herd of... Local llamas. Local llamas. Local llamas get loosened. That could be the headline. Could be a coincidence. SC-23 interrupted by local llama herd. But we're both from California, so we're not... You know, where is this Ohio place of which you speak? It's... We just fly over. It's on your way to Boston and New York. You'll see it out the window if you take a peek. Truth be told, I've been to Columbus twice in the last 30 days. You see? You're not so bad. You can come down south and hang out in the lab, which is one of the things like why I get so excited about hardware is because we've got a lab and we're touching these things and we're setting them up and we're going through the pains and much of it self-inflicted, I'll admit. We're not always perfect and clean and tidy in our approach, but it's very pragmatic in terms of getting these blade servers in. We've actually got four H100s in our lab right now, which is... I shouldn't say that when we're out of town. I was just going to say, what's your code? What's the address? Yeah, yeah, yeah. Kevin's manning the lab and we've got the intern on 24 hour watch. But we're doing all sorts of crazy stuff. We've got liquid nitrogen coming in tomorrow and we're going to do some overclocking world records, we hope. So there's so much out there, there's so much possibility. And I mean, you're talking too, I don't want to drive past this about what you do the work on. There's a lot of momentum around trying to do more AI work on CPUs. Exactly. And not just the GPUs. And of course, NVIDIA this week's talking a lot about Grace Hopper. We've seen systems from at least half a dozen vendors here that are either double grace systems or Grace Hopper super chips. I mean, there's so much potential. And there's to your point earlier about keeping busy and employed. I mean, I think there's going to be a lot of work. A lot of work needs to be done in exploring these solutions, looking past vendor claims and saying, what are these things really good at? And not necessarily siloing those, but isolating the benefits of ARM architectures, of DPUs, of all of these things and helping people sort out where the benefits are and to make reasonable investment decisions. My goodness, what a fantastic series of insights. I want to call out one last thing before I close the show. I wish more folks named things after strong bad ass women than numbers and letters as a configuration. Grace Hopper as a platform versus an H100. I feel like we can just get a little sexier with the naming. Brian Bealer, thank you so much for being here with us. Thanks for having me. Wonderful insights. And I look forward to having another chat with you again at some other NerdFest. David, always a pleasure to share the stage with you and thank all of you brilliant super compute fans for tuning in to our live coverage here from the mile high city. My name is Savannah Peterson. You're watching theCUBE, the leading source for emerging tech news.